Prediction of early responses to dupilumab treatment in patients with moderate-to-severe atopic dermatitis by a machine learning model integrating clinical characteristics and serum biomarkers.

AIM: The aim of the study was to evaluate the predictive value of baseline serum periostin and other T helper cell type 2 (Th2)-related biomarkers for early response to dupilumab in patients with moderate-to-severe atopic dermatitis (AD), and to develop a machine learning model integrating clinical and laboratory features. MATERIAL AND METHODS: A single-centre prospective cohort of 200 adults with moderate-to-severe AD treated with dupilumab (2020-2023) was used for model development, with an independent prospective cohort of 60 patients (2024) for external validation. Baseline clinical characteristics and serum biomarkers, including periostin, eosinophil count, total immunoglobulin E (IgE), and lactate dehydrogenase (LDH), were collected. Multivariate logistic regression and three machine learning models were constructed and evaluated using receiver operating characteristic analysis, calibration, decision curve analysis, and SHapley Additive exPlanations. RESULTS: In the modelling cohort, 62.5% of patients achieved EASI-75 at 16 weeks. Responders had significantly lower baseline periostin, eosinophil count, and IgE levels than non-responders (all p < 0.05). LASSO identified 10 key predictors, with periostin showing the strongest contribution. In multivariate logistic regression, periostin was the strongest independent predictor (OR = 0.34, p < 0.001) and improved discrimination. Among machine learning models, XGBoost performed best (AUC = 0.83). SHAP analysis confirmed periostin as the most influential feature. In external validation, the XGBoost model maintained good performance (AUC = 0.81; sensitivity = 0.917) with favourable calibration and clinical net benefit. CONCLUSIONS: Baseline serum periostin is a robust and independent biomarker for predicting early response to dupilumab in moderate-to-severe AD. A machine learning model integrating periostin with clinical and Th2-related biomarkers enables accurate and interpretable pre-treatment patient stratification.